search

choderalab / msm-pipeline

A pipeline for MSMs.
GNU Lesser General Public License v3.0
2 stars 5 forks source link

readme

Standardized MSM pipeline

Build Status

We have a large amount of simulation data from many related systems. Instead of selecting a different model-building approach for each system, we would like to select and automate a uniform MSM-building procedure, so that we can more easily compare MSMs.

Pipeline outline in one line

Trajectories --> discrete trajectories --> statistics

msm-pipeline /cbio/jclab/projects/fah/fah-data/munged-with-time/no-solvent/11400/run0-clone0.h5

Default project_name is 'abl', otherwise add project name in second argument:

msm-pipeline /cbio/jclab/projects/fah/fah-data/munged-with-time/no-solvent/11400/run0-clone0.h5 'abl'

Details

  1. Inputs:
    • List of filenames of munged hdf5 trajectories
      • These are the outputs of the FAHmunge pipeline
    • Lag-time between recorded snapshots
    • Lag-time for MSM estimation
    • Metastability lag-time
  2. Indexing:
    • Build an index of filename : nframes pairs.
    • Note that these files can be written to-- nframes can increase over time-- can't assume fixed file length, but can assume that the only file modifications will be to append to the end of the trajectory.
  3. Discretizing w.r.t. an appropriate metric
    • Two typical choices of metric:
      • Explicit-feature-based kinetic distance:
        • Compute features (angles, inter-residue distances, etc.)
        • Save computed features as a .npz archive
        • Use tICA to find a linear projection where Euclidean distance approximates diffusion distance
      • MinRMSD
    • Number of clusters. For now, we will fix this to be a large-ish number (1000), as we can recover from an overly fine discretization by "lumping" later.
    • Choice of clustering algorithm: for now, fix to minibatch k-means.
    • Save discrete trajectories as a .npz archive
  4. Statistics of discretized trajectories
    • Distribution of trajectory lengths
    • "Ergodic trimming" -- select the largest ergodic subspace, and estimate a Markov model
    • Test Markovianity
      • Implied timescales plots
      • Chapman-Kolmogorov tests among macrostates (macrostates will be defined below!)
    • "Lump" microstates into macrostates
      • Using eigenvalues of transition matrix

Major remaining challenges

  1. What do we do for "disconnected" models?
    • In cases where the observed dynamics are not ergodic, what should we do?
  2. What scheme gives the best model?